Hacker News new | ask | show | jobs
by sclarisse 1081 days ago
To support our next-gen machine learning system, we need a 10 exabyte storage array. It should host a system accessible over TCP/IP or Infiniband that can stream random-access 1MB blocks of data to 65,536 different computers, at continuous loads of 64 GB/second each computer, using a protocol of your choice or design. Correct for all data corruption and do not lose a single bit during the next 1,000 years of operation.

Yes, we can add more specs but these alone should be pretty daunting.

2 comments

What does this have to do with building software? This is something submitted on a form for hardware for a capex
If you think you’re gonna organize that many bits and not lose them for centuries at a time, you’re seriously underestimating the need for data replication and error correction algorithms, first of all.
Again, that's a hardware storage solution's problem. Adjust the parity level in the filesystem. Unless you're writing the requirements for building the software for a storage platform.
… yes, that is in fact the point
There are money, operations, and delivery time requirements, just off the top of my head, not expressed in that requirement that vastly changes the solution.
And the client is all over you/the company.